OctoInformBeta

OpenAI preprint, September 2026 · §4. The finite negative-type semimetric

Negative-type semimetrics and chart repair: Chart kernels, capped distances, chart triangle conditions, repair Gram representations, local cost, macro error, and positivity of the source constant

Chart triangle repair definitions and source metric parameters.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

Take an index set II, a set of chart labels Σ\Sigma, a chart map c:I→Σc : I \to \Sigma, and a map h:I→Hh : I \to H into a normed space HH. For points i,j∈Ii, j \in I in the same chart, the squared difference ∥h(i)−h(j)∥2\|h(i) - h(j)\|^2 gives a local distance, while pairs in different charts have no shared coordinates. How does one assemble these chartwise distances into Euclidean inner products for a negative-type semimetric up to a cutoff scale, and what explicit cost and error parameters govern the construction for a natural number mm?

Profile and Cost Functions

For a real number xx and a natural number nn, the inverse power p(x,n)p(x, n) is the real number 1/xn1/x^n.✦

The profile function ϕ:ℝ→ℝ\phi : \mathbb{R} \to \mathbb{R} is the function defined for each t∈ℝt \in \mathbb{R} by ϕ(t)=1−exp⁡(−1/(2t2))\phi(t) = 1 - \exp(-1/(2t^2)).✦

The profile constant c*c^* is the real number given by c*=2∫0∞ϕ(t)dt.c^* = 2 \int_0^\infty \phi(t)\,dt.(1)✦

The local cost Cloc(N,S,u,k,ε,a,b,R,γ)C_{\mathrm{loc}}(N, S, u, k, \varepsilon, a, b, R, \gamma) of natural numbers NN and SS and real numbers u,k,ε,a,b,R,γu, k, \varepsilon, a, b, R, \gamma is the real number given by Cloc(N,S,u,k,ε,a,b,R,γ)=2[(k+uγ(4k)+εSN(9k/ε)(kR/ε))log⁡(b/a)+εSN(2u)(4k)(1+(kR/ε)2+2(kR/ε))a].\begin{aligned} C_{\mathrm{loc}}(N, S, u, k, \varepsilon, a, b, R, \gamma) &= 2\Bigl[\bigl(k + u\gamma(4k) + \varepsilon S N (9k/\varepsilon)(kR/\varepsilon)\bigr)\log(b/a) \\ &\qquad + \frac{\varepsilon S N (2u)(4k) \bigl(1 + (kR/\varepsilon)^2 + 2(kR/\varepsilon)\bigr)}{a}\Bigr]. \end{aligned}(2)✦

The macro cost Cmac(N,S,u,k,ε,a,b,R,T,Rpos)C_{\mathrm{mac}}(N, S, u, k, \varepsilon, a, b, R, T, R_{\mathrm{pos}}) of natural numbers NN and SS and real numbers u,k,ε,a,b,R,T,Rposu, k, \varepsilon, a, b, R, T, R_{\mathrm{pos}} is the real number given by Cmac(N,S,u,k,ε,a,b,R,T,Rpos)=u(2a+(2T)2/b)+4εSNlog⁡(b/a)+4εSN((kR/ε)2+2(kR/ε))log⁡(b/a)+2c*uRpos,\begin{aligned} C_{\mathrm{mac}}(N, S, u, k, \varepsilon, a, b, R, T, R_{\mathrm{pos}}) &= u\bigl(2a + (2T)^2/b\bigr) + 4\varepsilon S N \log(b/a) \\ &\qquad + 4\varepsilon S N \bigl((kR/\varepsilon)^2 + 2(kR/\varepsilon)\bigr)\log(b/a) + 2c^* u R_{\mathrm{pos}}, \end{aligned}(3)✦

The Claim

Definition 1. Let II be an index set, let Σ\Sigma be a set of chart labels, let c:I→Σc : I \to \Sigma be a map, and let h:I→Hh : I \to H be a map into a normed space HH. For a real scale ℓ\ell, the chart kernel Kc,h(ℓ)K_{c,h}(\ell) is the I×II \times I real matrix defined for i,j∈Ii, j \in I by Kc,h(ℓ)i,j={exp⁡(−∥h(i)−h(j)∥2/ℓ)if c(i)=c(j),0if c(i)≠c(j).K_{c,h}(\ell)_{i,j} = \begin{cases} \exp\bigl(-\|h(i) - h(j)\|^2 / \ell\bigr) & \text{if } c(i) = c(j), \\ 0 & \text{if } c(i) \neq c(j). \end{cases}(4)✦ For a real cutoff rr, the capped distance dc,h,r:I×I→ℝd_{c,h,r} : I \times I \to \mathbb{R} is the function given for i,j∈Ii, j \in I by dc,h,r(i,j)={min⁡(r,∥h(i)−h(j)∥2)if c(i)=c(j),rif c(i)≠c(j).d_{c,h,r}(i, j) = \begin{cases} \min\bigl(r, \|h(i) - h(j)\|^2\bigr) & \text{if } c(i) = c(j), \\ r & \text{if } c(i) \neq c(j). \end{cases}(5)✦ A pair (c,h)(c, h) of a map c:I→Σc : I \to \Sigma and a map h:I→Hh : I \to H into a normed space HH satisfies the chart triangle inequality if for any i,j,k∈Ii, j, k \in I with c(i)=c(j)=c(k)c(i) = c(j) = c(k), ∥h(i)−h(k)∥2≤∥h(i)−h(j)∥2+∥h(j)−h(k)∥2.\|h(i) - h(k)\|^2 \le \|h(i) - h(j)\|^2 + \|h(j) - h(k)\|^2.(6)✦ Given a real cutoff rr, a map q:I→ℝIq : I \to \mathbb{R}^I into Euclidean space is a repair Gram representation for (c,h,r)(c, h, r) if for all i,j∈Ii, j \in I, ⟨q(i),q(j)⟩=4∫0rKc,h(ℓ)i,jdℓ,\langle q(i), q(j) \rangle = 4 \int_0^r K_{c,h}(\ell)_{i,j}\,d\ell,(7)✦

For a natural number mm, the local cost LL is the real number defined via (2) by L=Cloc(m6,m3,m,p(m,5),p(m,200),p(m,40),m40,p(m,1997),2/m),L = C_{\mathrm{loc}}\bigl(m^6, m^3, \sqrt{m}, p(m,5), p(m,200), p(m,40), m^{40}, p(m,1997), 2/\sqrt{m}\bigr),(8)✦ The macro error err\mathrm{err} of a natural number mm is the real number defined via (3) by err=Cmac(m6,m3,m,p(m,5),p(m,200),p(m,40),m40,p(m,1997),3m,0).\mathrm{err} = C_{\mathrm{mac}}\bigl(m^6, m^3, \sqrt{m}, p(m,5), p(m,200), p(m,40), m^{40}, p(m,1997), 3\sqrt{m}, 0\bigr).(9)✦

In terms of the profile constant c*c^* from (1), the source constant C0C_0 is the real number 9+2c*9 + 2c^*✦, and C0C_0 satisfies C0>0.C_0 > 0.(10)✦

What Is Checked

The chart kernel (4), capped distance (5), chart triangle inequality (6), and repair Gram representation (7) are stated over an index set II, a chart label set Σ\Sigma, a map c:I→Σc : I \to \Sigma, a map h:I→Hh : I \to H into a normed space HH, real parameters ℓ\ell and rr, and a map q:I→ℝIq : I \to \mathbb{R}^I into Euclidean space, whereas the local cost (8) and macro error (9) are stated over a natural number mm, and the source constant C0C_0 together with the inequality (10) carry no parameters.